-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmodel.py
131 lines (109 loc) · 7.8 KB
/
model.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
import tensorlayer as tl
import tensorflow as tf
class Net(object):
def work(self, imgs_ph, tags_ph, predict):
# Create the wrapper
self.predict = predict
self.tags_ph = tags_ph
# Others
self.loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=self.tags_ph, logits=self.predict))
self.optimize = tf.train.AdamOptimizer().minimize(self.loss)
self.accuracy = tf.reduce_mean(tf.cast( \
tf.equal(tf.argmax(self.predict, 1), tf.argmax(self.tags_ph, 1)), "float"), name='accuracy')
def zeroPadding(self, x):
self.zero_tensor = tf.zeros_like(x)
return tf.concat([x, self.zero_tensor], axis=3)
class CNN(Net):
def __init__(self, imgs_ph, tags_ph):
self.imgs_ph = imgs_ph
self.tags_ph = tags_ph
# 1st conv fix channel
self.network = tl.layers.InputLayer(self.imgs_ph, name='cnn_input')
self.network = tl.layers.Conv2d(self.network, n_filter=16, act = tf.nn.relu, name='cnn_conv1')
# 2 & 3 conv
self.network = tl.layers.Conv2d(self.network, n_filter=16, name='cnn_conv2')
self.network = tl.layers.Conv2d(self.network, n_filter=16, name='cnn_conv3')
self.network = tl.layers.BatchNormLayer(self.network, act = tf.nn.relu, name='cnn_bn_relu_1')
self.network = tl.layers.MaxPool2d(self.network, name='cnn_maxpool1')
# 4th conv fix channel
self.network = tl.layers.Conv2d(self.network, n_filter=32, act = tf.nn.relu, name='cnn_conv4')
# 5 & 6 conv
self.network = tl.layers.Conv2d(self.network, n_filter=32, name='cnn_conv5')
self.network = tl.layers.Conv2d(self.network, n_filter=32, name='cnn_conv6')
self.network = tl.layers.BatchNormLayer(self.network, act = tf.nn.relu, name='cnn_bn_relu_2')
self.network = tl.layers.MaxPool2d(self.network, name='cnn_maxpool2')
# Softmax
self.network = tl.layers.Conv2d(self.network, n_filter=64, act = tf.nn.relu, name='cnn_conv7')
self.network = tl.layers.FlattenLayer(self.network, name='cnn_flat')
self.network = tl.layers.DenseLayer(self.network, n_units = 10, act = tf.nn.softmax, name='cnn_fc')
self.predict = self.network.outputs
self.work(imgs_ph, tags_ph, self.predict)
class ResNet(Net):
def __init__(self, imgs_ph, tags_ph):
self.imgs_ph = imgs_ph
self.tags_ph = tags_ph
# Conv fix channel
self.network = tl.layers.InputLayer(self.imgs_ph, name='resnet_input')
self.input_layer = tl.layers.Conv2d(self.network, n_filter=16, act = tf.nn.relu, name='resnet_conv1')
# -----------------------------
# 1st general residual block
# -----------------------------
self.network = tl.layers.Conv2d(self.input_layer, n_filter=32, name='resnet_conv2')
self.network = tl.layers.Conv2d(self.network, n_filter=32, name='resnet_conv3')
padded_input = tl.layers.LambdaLayer(self.input_layer, self.zeroPadding, name='resnet_channel_padding_1')
self.network = tl.layers.ElementwiseLayer([self.network, padded_input], combine_fn = tf.add, name='resnet_add_1')
self.network = tl.layers.BatchNormLayer(self.network, act = tf.nn.relu, name='resnet_bn_relu_1')
self.input_layer = tl.layers.MaxPool2d(self.network, name='resnet_maxpool1')
# -----------------------------
# 2nd general residual block
# -----------------------------
self.network = tl.layers.Conv2d(self.input_layer, n_filter=64, name='resnet_conv5')
self.network = tl.layers.Conv2d(self.network, n_filter=64, name='resnet_conv6')
padded_input = tl.layers.LambdaLayer(self.input_layer, self.zeroPadding, name='resnet_channel_padding_2')
self.network = tl.layers.ElementwiseLayer([self.network, padded_input], combine_fn = tf.add, name='resnet_add_2')
self.network = tl.layers.BatchNormLayer(self.network, act = tf.nn.relu, name='resnet_bn_relu_2')
self.network = tl.layers.MaxPool2d(self.network, name='resnet_maxpool2')
# Softmax
self.network = tl.layers.Conv2d(self.network, n_filter=128, act = tf.nn.relu, name='resnet_conv7')
self.network = tl.layers.FlattenLayer(self.network, name='resnet_flat')
self.network = tl.layers.DenseLayer(self.network, n_units = 10, act = tf.nn.softmax, name='resnet_fc')
self.predict = self.network.outputs
self.work(imgs_ph, tags_ph, self.predict)
class RiR(Net):
def __init__(self, imgs_ph, tags_ph):
self.imgs_ph = imgs_ph
self.tags_ph = tags_ph
# Revise channel first
self.residual_input = tl.layers.InputLayer(imgs_ph, name ='rir_residual_input')
self.transient_input = tl.layers.InputLayer(imgs_ph, name ='rir_transient_input')
self.residual_stream = tl.layers.Conv2d(self.residual_input, act = tf.nn.relu, n_filter=16, name='rir_revise_conv1')
self.transient_stream = tl.layers.Conv2d(self.transient_input, act = tf.nn.relu, n_filter=16, name ='rir_revise_conv2')
# Add general residual blocks
self.residual_stream, self.transient_stream = self.residual_block(self.residual_stream, self.transient_stream, 32, 1)
self.residual_stream, self.transient_stream = self.residual_block(self.residual_stream, self.transient_stream, 64, 2)
# -----------------------------
# Softmax
# -----------------------------
self.network = tl.layers.ConcatLayer([self.residual_stream, self.transient_stream])
self.network = tl.layers.Conv2d(self.network, n_filter=128, act = tf.nn.relu, name='rir_conv7')
# Global average pooling
kernel_height = self.network.outputs.shape[1]
kernel_width = self.network.outputs.shape[2]
self.network = tl.layers.MaxPool2d(self.network, filter_size=(kernel_height, kernel_width))
self.network = tl.layers.FlattenLayer(self.network, name='rir_flat')
# Softmax
self.network = tl.layers.DenseLayer(self.network, n_units = 10, act = tf.nn.softmax, name='rir_fc')
self.predict = self.network.outputs
self.work(imgs_ph, tags_ph, self.predict)
def residual_block(self, residual_input, transient_input, n_filters, name):
self.residual_stram_main = tl.layers.Conv2d(residual_input, n_filter=n_filters, name='rir_general_residual_block_'+str(name)+'conv1')
self.residual_stram_extra = tl.layers.Conv2d(residual_input, n_filter=n_filters, name='rir_general_residual_block_'+str(name)+'conv2')
self.transient_stream_main = tl.layers.Conv2d(residual_input, n_filter=n_filters, name='rir_general_residual_block_'+str(name)+'conv3')
self.transient_stream_extra = tl.layers.Conv2d(residual_input, n_filter=n_filters, name='rir_general_residual_block_'+str(name)+'conv4')
self.residual_stram_main = tl.layers.ElementwiseLayer([self.residual_stram_main, self.transient_stream_extra], combine_fn = tf.add, name='rir_general_residual_block_'+str(name)+'add1')
padded_input = tl.layers.LambdaLayer(residual_input, self.zeroPadding, name='rir_general_residual_block_'+str(name)+'channel_padding')
self.residual_stram_main = tl.layers.ElementwiseLayer([self.residual_stram_main, padded_input], combine_fn = tf.add, name='rir_general_residual_block_'+str(name)+'add2')
self.transient_stream_main = tl.layers.ElementwiseLayer([self.transient_stream_main, self.residual_stram_extra], combine_fn = tf.add, name='rir_general_residual_block_'+str(name)+'add3')
self.residual_stram_main = tl.layers.BatchNormLayer(self.residual_stram_main, act = tf.nn.relu, name='rir_general_residual_block_'+str(name)+'_res_bn')
self.transient_stream_main = tl.layers.BatchNormLayer(self.transient_stream_main, act = tf.nn.relu, name='rir_general_residual_block_'+str(name)+'_tra_bn')
return self.residual_stram_main, self.transient_stream_main